The typical situation
Enquiries arrive by email and WhatsApp: a list of sections, grades and quantities, and a delivery location. Each one needs a stock check, the day's price list, freight and a reply. The sales team spends most of the day on enquiries that look almost identical.
The owner had seen an AI demo do this well on three examples and badly on the fourth, and wanted to know how often the fourth happens before letting it near a customer.
What the data looked like
A year of enquiries and the quotations actually sent, the stock system, the daily price list and the freight rate card. About 400 past enquiries, with the quotes that were sent, become the test set.
What we built
- An evaluation harness: 400 real past enquiries with the correct quote for each, scored automatically on items, quantities, price and freight.
- An agent that reads the enquiry, checks stock and prices, and drafts the quotation for a salesperson to approve.
- Rules that send anything unusual (a new customer, an out-of-stock grade, a quantity above a limit) straight to a person.
- A weekly report of approvals, edits and rejections, so the team can see where the agent is trusted and where it is not.
Typical outcome
| Measure | Before | After | Period |
|---|---|---|---|
| Score on the 400-enquiry test set before go-live | not measured | above the agreed 95% threshold | typical outcome |
| Routine quotes approved without edits | 0% | about 60% | typical, first two months |
| Time from enquiry to quote | half a day | under an hour | typical outcome |
What usually trips these projects up
The harness takes longer than the agent. Getting the correct answer for 400 old enquiries means sitting with the sales team, and that work is what makes the agent safe to switch on.